Superpixel-based Radial Basis Function Canonical Correlation Analysis for land cover segmentation

Hong, X. ORCID: https://orcid.org/0000-0002-6832-2298, Wei, H. ORCID: https://orcid.org/0000-0002-9664-5748 and Xia, J. (2025) Superpixel-based Radial Basis Function Canonical Correlation Analysis for land cover segmentation. International Journal of Remote Sensing, 46 (24). pp. 9411-9433. ISSN 1366-5901 doi: 10.1080/01431161.2025.2579806

Abstract/Summary

The radial basis functions (RBF) network is popularly used in many machine learning applications. Semantic segmentation of remotely sensed image which can be addressed via a clustering task, is essential in land cover applications. In this study, a novel semi-supervised spectral clustering method, referred to as SLIC-RBF-CCA, is introduced for land cover multi-band image segmentation. The proposed SLIC-RBF-CCA approach is composed of two steps: i) a Simple Linear Iterative Clustering (SLIC) algorithm is initially applied to a set of pseudo-RGB images obtained from the singular vectors of a flattened matrix of a multi-band image; ii) a novel superpixel-based Radial Basis Function Canonical Correlation Analysis (RBF-CCA) generates canonical variables which are used to achieve final image segmentation. Specifically, a superpixel-based radial basis function is defined as the first variable in the framework of Canonical Correlation Analysis, in which the RBF centers are obtained as the local mean from superpixel regions. The second variable of SLIC-RBF-CCA is based on a few labelled pixels. The associated canonical variables, related to the pixels of a full image, are then applied by a $k$-means clustering algorithm. The proposed approach can be interpreted as an example of multi-view machine learning with attention mechanism. Finally, the effectiveness of the proposed algorithm has been validated using experiments on several remotely sensed multi-band images, including two patches from TopoSys GmbH and three patches from City of Potsdam from the ISPRS, showing excellent performance with segmentation accuracy >85 % .

Item Type	Article
URI	https://centaur.reading.ac.uk/id/eprint/127443
Identification Number/DOI	10.1080/01431161.2025.2579806
Refereed	Yes
Divisions	Science > School of Mathematical, Physical and Computational Sciences > Department of Computer Science
Publisher	Taylor & Francis
Download/View statistics	View download statistics for this item